Fix DMA page lookup for IOMMU allocations

[xiaotianxt]

Summary

• add exanic_virt_to_page() to handle both direct-mapped and vmalloc-backed addresses
• use it at all five page lookups for memory returned by dma_alloc_coherent()

Root cause

With an IOMMU enabled, the DMA API can assemble a coherent allocation from separate pages and expose its CPU mapping through vmap(). Such an address is in vmalloc space. On x86, virt_to_page() is only valid when virt_addr_valid() is true, so applying it to that mapping produces an invalid struct page * and can fault in page_count() or vm_insert_page().

The helper follows the same split used by the Linux IOMMU DMA implementation when releasing coherent allocations: vmalloc_to_page() for vmalloc addresses and virt_to_page() otherwise.

This addresses the IOMMU-dependent kernel faults reported in #94 and #95. A matching vm_insert_page() fault with IOMMU enabled was also reported in #70.

Impact

Avoid kernel page faults while mapping ExaNIC DMA buffers and while checking their mapping reference counts. Non-IOMMU/direct-mapped allocations keep the existing behavior.

Validation

• built exanic.ko against Debian 12 x86_64 Linux 6.1.180 headers
• built exanic.ko against Rocky Linux 9 x86_64 Linux 5.14.0-687.36.1 headers
• verified that all five DMA page lookups route through the helper

Runtime validation still requires an ExaNIC-equipped Linux host with IOMMU enabled.